EMELD Conference July 2004, Detroit Archiving Language Resource - - PowerPoint PPT Presentation

http://www.spectrum.uni-bielefeld.de/modelex/

1

Archiving Language Resource Objects in XML: Experiences with TAMINO

Dafydd Gibbon, Thorsten Trippel, Ben Hell Universität Bielefeld, Europe

{gibbon|ttrippel|ben}@spectrum.uni-bielefeld.de

EMELD Conference July 2004, Detroit

http://www.spectrum.uni-bielefeld.de/modelex/

2

Overview

• Archiving in XML
• Language Resources
• Getting abstract: types of Resource Object
• Abstract Resource Object implementation in XML
• Getting Practical: the ModeLex application
• Using an XML database: TAMINO

Procedure - Database creation - Corpus data stored in the file system - Using a DBMS for storing Resource Objects - Selected use cases - Querying - Signal processing

• Conclusion: evaluation and further work

http://www.spectrum.uni-bielefeld.de/modelex/

3

Archiving in XML

Background: projects

Ega (2001) ModeLex (2001-...) ABUILD (2002-...) LLSTI (2003-...)

Goal: specifying a DBMS for Resource Object storage Resource Objects:

General Resource Object (GRO, linguistic data type) Specific Resource Object (SRO, instance of GRO) Abstract Resource Object (ARO, abstract data structure) Implementational Resource Object (IRO, PL/KRL data structure)

http://www.spectrum.uni-bielefeld.de/modelex/

4

Written texts, dialogue transcriptions Annotations

time-stamped transcription marked up written text & transcription

Signal recordings

audio, video, laryngograph (electroglottograph), airflow, ...

Lexical information Multimodal resource search:

structuring with XML storing accessing updating

Language Resources

http://www.spectrum.uni-bielefeld.de/modelex/

5

Getting abstract: types of Resource Object

General Resource Object (GRO, linguistic data type) Specific Resource Object (SRO, instance of GRO) Abstract Resource Object (ARO, abstract data structure)

strings - string sequences - structures over strings - lists - tables - DAGs - CGs - numbers - ...

Implementational Resource Object (IRO, PL/KL data structure) TREES: typically for constituent structures and taxonomies TABLES: typically for lexica and paradigm tables DAGs: typically for (almost) anything ☺

http://www.spectrum.uni-bielefeld.de/modelex/

6

XML abstract syntax defined as recursive ternary relation...

OBJECT = string OBJECT = {x: x = <typename, AVS, OBJECT+>}

... can only define tree structures: an bn (Type 2, CF L) Not defined in XML syntax: For embedded tables further constraints necessary... ... general indexing needed: an bn cn (Type 1, CS L subset) For general graphs, networks, semantic extension needed: pointer structures permit extension beyond tree structures.

Thus: access tools must be more powerful than XML syntax requires.

Abstract Resource Objects and XML IROs

http://www.spectrum.uni-bielefeld.de/modelex/

7

Getting practical: the ModeLex application

Subcorpus Annotation layer Annotation segment Typisation (Metadata) Segment in context Signal in context

Corpus layers and metadata layers: Search application:

http://www.spectrum.uni-bielefeld.de/modelex/

8

Preliminaries: XML format normalisation

Corpus format: depends on application (WAV; praat, esps-waves+, TASX, ...) Normalization: XML format Preservation of all bits of information from source metadata timestamps technical information Time Aligned Signal eXchange format (TASX) Grammar normalization: DTD to XSchema conversion

http://www.spectrum.uni-bielefeld.de/modelex/

9

Use case: Multimodal concordance

Functional requirements specification: Input: <searchkey, <recording, annotation>> Output: subset of <recording, anotation>

• matching search key + context-tier
• corresponding to output format filters

(tiers, length, signal transformation)

http://www.spectrum.uni-bielefeld.de/modelex/

10

Design: signal concordance

http://www.spectrum.uni-bielefeld.de/modelex/

11

Implementation: TAMINO XML DBMS - 1

Options:

• 1. data on file system:

command line access easy to manipulate selection complex performance with large repositories

• 2. storage in TAMINO

create DB (Tamino Software Management Hub) create "collection" (Tamino: Schema Editor) insert schema (Tamino: Schema Editor) insert document instances (Tamino: Schema Editor, Tamino Interactive Interface,...)

http://www.spectrum.uni-bielefeld.de/modelex/

12

Implementation: TAMINO XML DBMS - 2

Traditional tools: file system + ad hoc tools XML command technologies: filesystem based XQuery saxon XQuery tool /Java Library exquisit: GUI for saxon Tamino based tools: Tamino interactive interface, webinterface Tamino XQuery (Windows Application) Tamino Java API Perl API: any Perl program, e.g. browser based GUIs

http://www.spectrum.uni-bielefeld.de/modelex/

13

Access: based on XQuery unit selection using metadata AND annotation segment key context selection on same tier OR parallel tiers based on time interval → XQuery arithmetic in Tamino sibling access → available in saxon, not in Tamino

Implementation: TAMINO XML DBMS - 3

http://www.spectrum.uni-bielefeld.de/modelex/

14

Audio: Selected interval based on time stamps Further analysis possible if lossless compression files: spectrogram, oscillogram, formant analysis, ... Fast (almost real time): praat scripting + sox

Gibbon and Trippel 2001: Portable Audio Concordance System. TR-UBI

Video: Audio in principle as above Granularity: frame based, not sample-based technical restrictions: keyframe rate Time consuming: no real time processing

Implementation: TAMINO XML DBMS - 4

http://www.spectrum.uni-bielefeld.de/modelex/

15

Summary:

Proof of concept for TASX audio corpus

Tamino, Perl Audio signal processing: PAX modules, based on Praat

XQuery selection: corpus - subcorpus - layer - segment To do:

GUI not fail-safe (fails if metadata incomplete) Inconsistency potential in file storage of signal recordings Optimisation of XQuery vs. XSLT for formatting

http://www.spectrum.uni-bielefeld.de/modelex/implementation/concordance.html

Conclusion: evaluation and further work